Fixed
- The ROOT_EVAL coverage gate no longer finalizes a goal when its Eval step merely executed successfully. It now finalizes only on a structured
complete=true coverage_verdict, so an Eval that returned an "incomplete" verdict as prose (no continuation subtasks) forces a bounded re-audit instead of silently closing the goal.
- Roll back consumed relay-inbox entries when a malformed
pending_clarification_answer raises during parsing, so the user's answer is re-enqueued at the inbox head instead of being permanently lost.
- Emit
clarification_requested early in the execute node (before routing to await_clarification) so the event reliably reaches the broadcast pipeline; also emit on resume turns where the await node skips its own emit.
- Bump
soothe-nano floor to 1.3.2.
Changed
- Eval steps emit a binding structured
coverage_verdict tool call (complete or remaining subtasks) instead of a free-prose verdict; decompose_task is no longer injected on Eval threads. The verdict rides the existing step execution.outcome dict, with no StepNode schema or persistence migration.
Compare with previous version
